CN105787458A - Infrared behavior identification method based on adaptive fusion of artificial design feature and depth learning feature - Google Patents

Infrared behavior identification method based on adaptive fusion of artificial design feature and depth learning feature Download PDF

Info

Publication number
CN105787458A
CN105787458A CN201610139450.8A CN201610139450A CN105787458A CN 105787458 A CN105787458 A CN 105787458A CN 201610139450 A CN201610139450 A CN 201610139450A CN 105787458 A CN105787458 A CN 105787458A
Authority
CN
China
Prior art keywords
feature
features
optical flow
network
weight
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201610139450.8A
Other languages
Chinese (zh)
Other versions
CN105787458B (en
Inventor
高陈强
吕静
杜银和
汪澜
刘江
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN201610139450.8A priority Critical patent/CN105787458B/en
Publication of CN105787458A publication Critical patent/CN105787458A/en
Application granted granted Critical
Publication of CN105787458B publication Critical patent/CN105787458B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20—Movements or behaviour, e.g. gesture recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Psychiatry (AREA)
  • Social Psychology (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

本发明涉及一种基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，包括以下步骤：S1：通过人工设计特征模块，对原始视频进行改进的密集轨迹特征提取；S2：对提取到的人工设计特征进行特征编码；S3：通过CNN特征模块，对原始视频图像序列，利用一种变分光流算法提取光流信息，得到对应光流图像序列；S4：利用卷积神经网络，对步骤S3中得到的光流图序列提取CNN特征；S5：将数据集分为训练集和测试集；对训练集数据，通过权重优化网络学习权重，利用学习到的权重将CNN特征分类网络和人工设计特征分类网络的概率输出进行加权融合，通过对比识别结果得出最优权重，并将其应用于测试集数据分类。本方法创新了特征融合方式，提高了红外视频中行为识别的可靠性，对后续视频分析工作有重大意义。

The invention relates to an infrared behavior recognition method based on adaptive fusion of artificially designed features and deep learning features, comprising the following steps: S1: extracting dense trajectory features from the original video through artificially designed feature modules; S2: extracting the extracted features S3: Using the CNN feature module, use a variational optical flow algorithm to extract the optical flow information of the original video image sequence, and obtain the corresponding optical flow image sequence; S4: Using the convolutional neural network, the step The optical flow map sequence obtained in S3 extracts CNN features; S5: Divide the data set into a training set and a test set; for the training set data, optimize the network learning weights through weights, and use the learned weights to classify CNN features into network and manual design The probability output of the feature classification network is weighted and fused, and the optimal weight is obtained by comparing the recognition results, and it is applied to the classification of the test set data. This method innovates the way of feature fusion, improves the reliability of behavior recognition in infrared video, and is of great significance to the follow-up video analysis work.

Description

Based on the infrared behavior recognition methods that engineer's feature and degree of deep learning characteristic self adaptation merge
Technical field
The invention belongs to image procossing and technical field of computer vision, relate to a kind of infrared behavior recognition methods merged based on engineer's feature and degree of deep learning characteristic self adaptation.
Background technology
In recent years, image and the Activity recognition in video become a vital task of computer vision field, and the work such as video monitoring, Video Information Retrieval Techniques, man-machine interaction are all of great significance by the Activity recognition in video.Along with various actions recognizer constantly refreshes the recognition accuracy of each public data collection, the Activity recognition task in video has made great progress.But current most data set is all based on visible light video, the Activity recognition based on infrared video works relatively fewer.
The Activity recognition algorithm of current main flow relates generally to two kinds of description: engineer's Feature Descriptor and description obtained by degree of depth study.
Engineer's feature and typical local description, such as space-time interest points feature (Spatial-TemporalInterestPoint, STIP), histograms of oriented gradients (HistogramofOrientedGradient, HOG), light stream direction histogram (HistogramofOpticalFlow, HOF), intensive track characteristic (DenseTrajectory, DT) etc., it is based on movable information between the texture information of image, vision shape and different frame etc. different behaviors is classified and identifies；Owing to intensive track characteristic comprises the abundant sub-information of description such as HOG, HOF, MBH (MotionBoundaryHistory), become engineer's feature that current recognition accuracy is higher；Lifting along with computing power, the CNN feature extracted by convolutional neural networks is used as image, Activity recognition in video becomes popular research direction in recent years, convolutional neural networks achieves the profound excavation to pictorial information, effective extraction has the information of identification, the 3D convolutional neural networks having been proposed that at present, degree of depth convolutional neural networks, in the models such as dual pathways convolutional neural networks, effect is a dual pathways convolutional neural networks being made up of temporal information passage and spatial information passage most preferably, each more challenging data set all achieves good experiment effect.
But the Activity recognition research work being currently based on infrared video is relatively fewer, in video monitoring works, if running into the relatively low vile weather of the visibility such as rain, mist or nighttime conditions descends, it is seen that light video monitoring just loses meaning.Therefore, infrared video Activity recognition has critically important practical value, is badly in need of a kind of effective infrared video Activity recognition algorithm at present.
Summary of the invention
In view of this, it is an object of the invention to provide a kind of infrared behavior recognition methods merged based on engineer's feature and degree of deep learning characteristic self adaptation, the method takes full advantage of feature and the advantage of infrared image, engineer's feature is improved, and by two sorter networks and a weight learning function, it is weighted the probability output of engineer's tagsort network and degree of deep learning characteristic sorter network merging, the accuracy rate being effectively improved in infrared video Activity recognition.
For reaching above-mentioned purpose, the present invention provides following technical scheme:
A kind of infrared behavior recognition methods merged based on engineer's feature and degree of deep learning characteristic self adaptation, the method comprises the following steps:
S1: by engineer's characteristic module, the intensive track characteristic that original video is improved extracts；
S2: the engineer's feature extracted in step S1 is carried out feature coding；
S3: by CNN characteristic module, to original video image sequence, utilizes a kind of variation optical flow algorithm to extract Optic flow information, obtains corresponding light flow image sequences, as the input of convolutional neural networks；
S4: utilize convolutional neural networks, extracts CNN feature to the light stream graphic sequence obtained in step S3；
S5: data set is divided into training set and test set；To training set data, by weight optimization e-learning weight, utilize study to weight be weighted the probability output of CNN tagsort network and engineer's tagsort network merging, draw optimal weights by contrasting recognition result, and be applied to the classification of test set data.
Further, in step sl, described to the engineer's feature after original video extraction improvement, specifically include: first each frame infrared image is carried out point of interest intensive sampling, then between consecutive image sequence, follow the trail of point of interest, describe addition track correspondence Pictures location gray value weight in son at the track obtained, the track after weighting is described the son intensive track characteristic as improvement.
Further, in step s 2, described is encoded intensive track characteristic after the improvement extracted, and coding carries out in the following ways:
1): Bag-of-Words (BOW)
It is step by step rapid that this feature coding method includes two:
11) visual dictionary is generated: in training set data, extract feature, by clustering algorithm, the feature extracted is clustered；Cluster centre can be regarded as the visual vocabulary in dictionary, and all visual vocabularies form a visual dictionary；
12) with the lexical representation image in dictionary: each feature in image is mapped on certain word of visual dictionary, then add up each visual word occurrence number on an image, image table can be shown as the histogram vectors that a dimension is fixing；
2): FisherVector (FV)
FisherVector coding is a kind of coded system based on Fisher core principle, gauss hybrid models (GMM) is trained by the method for maximal possibility estimation first with training sample, then utilizing gauss hybrid models that the primitive character (such as Dense-Traj) extracted from sample is modeled, the primitive character of sample is encoded into the Fisher vector being easy to study and tolerance by the model parameter that recycling generates；
3): VectorofLocallyAggregatedDescriptors (VLAD)
First the feature extracted in training set is clustered by VLAD coding, obtain code book, then calculate the gap of each word in primitive character and code book, and these differences are added up, finally the accumulative difference of all words is carried out being spliced to form a new vector to represent image.
Further, in step s3, described utilization one variation optical flow algorithm extracts Optic flow information, obtains corresponding light flow image sequences, specifically includes:
S31: when the constant hypothesis of brightness, constant gradient hypothesis and space-time smoothness constraint continuously, it is proposed to the energy functional of improvement；
S32: derive the Euler-Lagrange equation of its correspondence according to energy function, then utilizes Gauss-Seidel or SOR method to obtain light stream vector, obtains corresponding light flow graph.
Further, in step s 4, to the light flow image sequences obtained, utilize convolutional neural networks, extract the output of full articulamentum as CNN feature；CNN adopts multitiered network structure, convolutional layer and sampling layer to be alternately present, and network output layer is the full connected mode of feedforward network.
Further, in step s 5, described data set is divided into training set and test set；Utilize training set data to pass through the study of self adaptation Fusion Module and obtain optimal weights, and optimal weights is applied to test set data classifies, specifically include:
S51: utilize training set data to learn optimal weights:
Self adaptation Fusion Module includes two sorter networks comprising full articulamentum and softmax grader layer and a single node logical function；The input of two sorter networks is engineer's feature and the CNN feature of training set data respectively, and two kinds of features obtain respective probability output P1 and P2 after being input to map network, utilizes single node logical function to calculate the weight Q of correspondence simultaneously；The final probability output of two sorter networks being weighted according to specified weight is merged, and utilizes error backpropagation algorithm to judge identification error, update gradient, it is judged that and export optimal weights；
S52: optimal weights is applied to test set data and classifies:
Engineer's feature test set data extracted and CNN feature are input to corresponding sorter network, obtain corresponding probability output, utilize the probability output of two sorter networks of optimal weights Weighted Fusion obtained in S51, obtain the recognition result of test set data.
The beneficial effects of the present invention is: the method for the invention highlights the infrared video advantage compared with visible light video, Traditional Man design feature and the degree of deep learning characteristic recently quickly grown are conjointly employed in Activity recognition task, innovate Feature Fusion mode, improve the reliability of Activity recognition in infrared video, subsequent video analysis work is of great importance.
Accompanying drawing explanation
In order to make the purpose of the present invention, technical scheme and beneficial effect clearly, the present invention provides drawings described below to illustrate:
Fig. 1 is the dual pathways infrared behavior recognizer schematic diagram of the present invention；
Fig. 2 is that the intensive track characteristic improved extracts schematic diagram；
Fig. 3 is light stream convolutional neural networks feature extraction schematic diagram；
Fig. 4 is self-adaptive features Fusion Model schematic diagram.
Detailed description of the invention
Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.
In the present invention, engineer's characteristic module, the intensive track characteristic after original video is improved extracts, and the feature extracted is encoded；The intensive track characteristic improved is add gray value weight information in description of original intensive track, the space time information of sequence of video images of the intensive track characteristic major embodiment after improvement, highlights the foreground moving information of sequence of pictures；CNN characteristic module, original Infrared video image sequence utilize a kind of variation optical flow algorithm extract Optic flow information, form light flow image sequences, using the light flow graph that extracts as the input of convolutional neural networks, take the feature of the full articulamentum of convolutional neural networks as video CNN Feature Descriptor；First data set is divided into training set and test set by self adaptation Fusion Module；To training set data, weight is learnt by weight learning network, utilize study to weight be weighted the probability output of CNN tagsort network and engineer's tagsort network merging, draw optimal weights by contrasting recognition result, and optimal weights be applied to the classification of test set data.
Fig. 1 is the dual pathways infrared behavior recognizer schematic diagram of the present invention, as it can be seen, the method for the invention specifically includes following steps:
Step S1: to original video extract improve after intensive track characteristic, its particular flow sheet as shown in Figure 2:
S101: each frame infrared image of original video is carried out point of interest intensive sampling；
S102: following the trail of point of interest between consecutive image sequence, the pixel that between two continuous frames, displacement is only small and excessive is all rejected, and finally gives the point of interest track that can characterize action；
S103: describe addition track correspondence Pictures location gray value weight in son at the track obtained, describes the son intensive track characteristic as improvement using the track after weighting.
Original video is resolution be 293 × 256 infrared data collection, totally 12 type of action, each type of action has 50 video samples.Intensive track characteristic is engineer's feature that in current image Classification and Identification task, accuracy rate is higher, but the data set being mostly used in test is all visible data collection.Owing to infrared image can reflect the infrared emanation of each target in image, heat radiation is more strong, gray value is more big, making there is obvious contrast between target and background, based on this feature of infrared image, we add gray value weight on the basis of original intensive track characteristic, the track making gray value high has bigger weight, the track that gray value is less distributes less weight, improves original intensive track characteristic, highlights the advantage of infrared image Activity recognition.
Step S2: the engineer's feature extracted is carried out FisherVector coding.
Step S3: utilize a kind of variation optical flow algorithm that original video image sequence extracts Optic flow information, form light flow image sequences.The data item of this algorithm energy functional is made up of the constant hypothesis of brightness and the constant hypothesis of gradient, adds discontinuous space-time smoothness constraint, has continuous preferably and rotational invariance, calculates speed fast, and precision is high.
Step S4: the picture of the light flow image sequences that step S3 is obtained carries out pretreatment, it is 227 × 227 that dimension of picture adjusts resolution, and it can be used as the input of convolutional neural networks, this convolutional neural networks is made up of five convolutional layers and three full articulamentums, finally taking the feature of second convolutional layer as feature representation, flow chart is as shown in Figure 3.
Convolutional neural networks (ConvolutionalNeuralNet, CNN) it is a kind of Multi-layered Feedforward Networks, convolutional layer and sampling layer as network intermediate layer are alternately present, and network is output as the full connected mode of feedforward network, and the dimension of output layer is the classification number in classification task.Convolutional neural networks directly can automatically learn individual features from original input data, saves the characteristic Design process that general algorithm needs, it is possible to learn and find more effective feature.
Step S5: training set data utilizes self adaptation Fusion Module to learn optimal weights, and is applied to the classification of test set data.The schematic diagram of self adaptation Fusion Model is as shown in Figure 4:
S501: data set is divided into training set and test set；
S502: utilize training set sample data to learn optimal weights: self adaptation Fusion Model includes two sorter networks comprising full articulamentum and softmax grader layer and a single node logical function；The input of two sorter networks is engineer's feature and CNN feature that training set data is extracted respectively, and two kinds of features obtain respective probability output P1 and P2 after being input to map network, single node logical function calculates corresponding weight Q simultaneously；The final probability output of two sorter networks being weighted according to specified weight is merged, and utilizes error backpropagation algorithm to judge identification error, update gradient, it is judged that and export optimal weights；
S503: optimal weights is applied to test set data and classifies: engineer's feature test set sample data extracted and CNN feature are input to corresponding sorter network, obtain corresponding probability output, utilize optimal weights proportion weighted to merge the probability output of two sorter networks, obtain test set data recognition result.
The main purpose of error back propagation (ErrorBackpropagation, BP) algorithm is by by output error anti-pass, giving all unit of each layer by error distribution, thus obtaining the error signal of each layer unit, and then revises the weights of each unit.The learning process of BP algorithm is made up of two processes of back propagation of the forward-propagating of signal Yu error.During forward-propagating, input sample is incoming from input layer, after each hidden layer successively processes, is transmitted to output layer.If the actual output of output layer and desired output are not inconsistent, then proceed to the back-propagation phase of error.Error-duration model be by output error with some form by each hidden layer to input layer successively anti-pass, and error distribution is given all unit of each layer, thus obtaining the error signal of each layer unit, this error signal is namely as the foundation of correction unit weights.In this module, the process that this weights gone round and begun again constantly adjust just makes to finally give optimal weights.
What finally illustrate is, preferred embodiment above is only in order to illustrate technical scheme and unrestricted, although the present invention being described in detail by above preferred embodiment, but skilled artisan would appreciate that, in the form and details it can be made various change, without departing from claims of the present invention limited range.

Claims (6)

1.一种基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，其特征在于：该方法包括以下步骤：1. An infrared behavior recognition method based on artificial design features and deep learning feature adaptive fusion, is characterized in that: the method comprises the following steps: S1：通过人工设计特征模块，对原始视频进行改进的密集轨迹特征提取；S1: Through the manual design of the feature module, the improved dense trajectory feature extraction is performed on the original video; S2：对步骤S1中提取到的人工设计特征进行特征编码；S2: performing feature encoding on the artificially designed features extracted in step S1; S3：通过CNN特征模块，对原始视频图像序列，利用一种变分光流算法提取光流信息，得到对应光流图像序列，作为卷积神经网络的输入；S3: Through the CNN feature module, use a variational optical flow algorithm to extract the optical flow information for the original video image sequence, and obtain the corresponding optical flow image sequence as the input of the convolutional neural network; S4：利用卷积神经网络，对步骤S3中得到的光流图序列提取CNN特征；S4: using a convolutional neural network to extract CNN features from the optical flow graph sequence obtained in step S3; S5：将数据集分为训练集和测试集；对训练集数据，通过权重优化网络学习权重，利用学习到的权重将CNN特征分类网络和人工设计特征分类网络的概率输出进行加权融合，通过对比识别结果得出最优权重，并将其应用于测试集数据分类。S5: Divide the data set into a training set and a test set; for the training set data, optimize the network learning weight through the weight, and use the learned weight to weight the probability output of the CNN feature classification network and the artificially designed feature classification network. Through comparison The recognition results yield the optimal weights, which are applied to the test set data classification. 2.根据权利要求1所述的基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，其特征在于：在步骤S1中，所述对原始视频提取改进后的人工设计特征，具体包括：首先对每一帧红外图像进行兴趣点密集采样，然后在连续图像序列之间追踪兴趣点，在得到的轨迹描述子中加入轨迹对应图片位置灰度值权重，将加权后的轨迹描述子作为改进的密集轨迹特征。2. The infrared behavior recognition method based on artificial design features and deep learning feature adaptive fusion according to claim 1, characterized in that: in step S1, the improved artificial design features extracted from the original video, specifically include : First, densely sample interest points for each frame of infrared image, then track the interest points between consecutive image sequences, add the gray value weight of the trajectory corresponding to the image position in the obtained trajectory descriptor, and use the weighted trajectory descriptor as Improved dense trajectory features. 3.根据权利要求1所述的基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，其特征在于：在步骤S2中，所述的对提取的改进后密集轨迹特征进行编码，编码采用以下方式进行：3. The infrared behavior recognition method based on artificial design features and deep learning feature adaptive fusion according to claim 1, characterized in that: in step S2, the described improved dense trajectory feature after extraction is encoded, and the encoded Proceed in the following ways: 1)：Bag-of-Words(BOW)1): Bag-of-Words (BOW) 该特征编码方法包括两步步骤：This feature encoding method consists of two steps: 11)生成视觉词典：在训练集数据上提取特征，通过聚类算法对提取到的特征进行聚类；聚类中心可以看作是词典中的视觉词汇，所有视觉词汇形成一个视觉词典；11) Generate a visual dictionary: extract features from the training set data, and cluster the extracted features through a clustering algorithm; the cluster center can be regarded as the visual vocabulary in the dictionary, and all visual vocabulary forms a visual dictionary; 12)用词典中的词汇表示图像：图像中的每个特征都被映射到视觉词典的某个词上，然后统计每个视觉词在一张图像上的出现次数，即可将图像表示成一个维数固定的直方图向量；12) Represent the image with the words in the dictionary: each feature in the image is mapped to a word in the visual dictionary, and then the number of occurrences of each visual word on an image is counted, and the image can be represented as a a histogram vector of fixed dimensionality; 2)：FisherVector(FV)2): Fisher Vector (FV) FisherVector编码是一种基于Fisher核原理的编码方式，首先利用训练样本通过最大似然估计的方法训练出高斯混合模型(GMM)，然后利用高斯混合模型对从样本中提取的原始特征(如Dense-Traj)进行建模，再利用生成的模型参数将样本的原始特征编码成便于学习和度量的Fisher向量；FisherVector coding is a coding method based on the Fisher kernel principle. First, the Gaussian mixture model (GMM) is trained by using the training sample through the method of maximum likelihood estimation, and then the original features extracted from the sample (such as Dense- Traj) to model, and then use the generated model parameters to encode the original features of the sample into Fisher vectors that are easy to learn and measure; 3)：VectorofLocallyAggregatedDescriptors(VLAD)3): Vector of Locally Aggregated Descriptors (VLAD) VLAD编码首先对训练集上提取的特征进行聚类，得到码本，然后计算原始特征和码本中每个词的差距，并将这些差值进行累计，最后把所有词的累计差值进行拼接形成一个新的向量来表示图像。VLAD encoding first clusters the features extracted on the training set to obtain the codebook, then calculates the difference between the original features and each word in the codebook, and accumulates these differences, and finally splices the accumulated differences of all words Form a new vector to represent the image. 4.根据权利要求1所述的基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，其特征在于：在步骤S3中，所述的利用一种变分光流算法提取光流信息，得到对应光流图像序列，具体包括：4. The infrared behavior recognition method based on artificial design features and deep learning feature adaptive fusion according to claim 1, characterized in that: in step S3, the described utilization of a variational optical flow algorithm to extract optical flow information, Obtain the corresponding optical flow image sequence, including: S31：在亮度不变假设、梯度恒定假设和连续时空平滑约束条件下，提出改进的能量泛函；S31: Under the assumption of constant brightness, constant gradient and continuous space-time smooth constraints, an improved energy functional is proposed; S32：根据能量函数推导出其对应的Euler-Lagrange方程，然后利用Gauss-Seidel或SOR方法求出光流矢量，得到对应光流图。S32: Deduce the corresponding Euler-Lagrange equation according to the energy function, and then use the Gauss-Seidel or SOR method to calculate the optical flow vector to obtain the corresponding optical flow diagram. 5.根据权利要求1所述的基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，其特征在于：在步骤S4中，对得到的光流图像序列，利用卷积神经网络，提取全连接层的输出作为CNN特征；CNN采用多层网络结构，卷积层和抽样层交替出现，网络输出层为前馈网络全连接方式。5. The infrared behavior recognition method based on artificial design features and deep learning feature adaptive fusion according to claim 1, characterized in that: in step S4, the obtained optical flow image sequence is extracted using a convolutional neural network The output of the fully connected layer is used as the CNN feature; CNN adopts a multi-layer network structure, the convolutional layer and the sampling layer appear alternately, and the network output layer is a feedforward network fully connected. 6.根据权利要求1所述的基于人工设计特征和深度学习特征自适应融合的红外行为识别方法，其特征在于：在步骤S5中，所述将数据集分为训练集和测试集；利用训练集数据通过自适应融合模块学习得到最优权重，并将最优权重应用到测试集数据进行分类，具体包括：6. the infrared behavior recognition method based on manual design feature and deep learning feature adaptive fusion according to claim 1, is characterized in that: in step S5, described dataset is divided into training set and test set; The set data is learned through the adaptive fusion module to obtain the optimal weight, and the optimal weight is applied to the test set data for classification, including: S51：利用训练集数据学习出最优权重：S51: Use the training set data to learn the optimal weight: 自适应融合模块包括两个包含全连接层和softmax分类器层的分类网络和一个单节点逻辑函数；两个分类网络的输入分别是训练集数据的人工设计特征和CNN特征，两种特征输入到对应网络之后得到各自的概率输出P1和P2，同时利用单节点逻辑函数计算出对应的权重Q；最终根据指定权重对两个分类网络的概率输出进行加权融合，利用误差反向传播算法判断识别误差、更新梯度，判断并输出最优权重；The adaptive fusion module includes two classification networks including a fully connected layer and a softmax classifier layer and a single-node logic function; the inputs of the two classification networks are the artificially designed features and CNN features of the training set data, and the two features are input to After the corresponding network, the respective probability outputs P1 and P2 are obtained, and the corresponding weight Q is calculated by using the single-node logic function; finally, the probability outputs of the two classification networks are weighted and fused according to the specified weight, and the error back propagation algorithm is used to judge the recognition error , update the gradient, judge and output the optimal weight; S52：将最优权重应用到测试集数据进行分类：S52: Apply the optimal weight to the test set data for classification: 将测试集数据提取到的人工设计特征和CNN特征输入到对应分类网络，得到对应概率输出，利用S51中得到的最优权重加权融合两个分类网络的概率输出，得到测试集数据的识别结果。Input the artificially designed features and CNN features extracted from the test set data into the corresponding classification network to obtain the corresponding probability output, and use the optimal weight obtained in S51 to weight the probability outputs of the two classification networks to obtain the recognition result of the test set data.
CN201610139450.8A 2016-03-11 2016-03-11 The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature Active CN105787458B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610139450.8A CN105787458B (en) 2016-03-11 2016-03-11 The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610139450.8A CN105787458B (en) 2016-03-11 2016-03-11 The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature

Publications (2)

Publication Number Publication Date
CN105787458A true CN105787458A (en) 2016-07-20
CN105787458B CN105787458B (en) 2019-01-04

Family

ID=56392545

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610139450.8A Active CN105787458B (en) 2016-03-11 2016-03-11 The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature

Country Status (1)

Country Link
CN (1) CN105787458B (en)

Cited By (50)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106355210A (en) * 2016-09-14 2017-01-25 华北电力大学（保定） Method for expressing infrared image features of insulators on basis of depth neuron response modes
CN106446937A (en) * 2016-09-08 2017-02-22 天津大学 Multi-convolution identifying system for AER image sensor
CN106709447A (en) * 2016-12-21 2017-05-24 华南理工大学 Abnormal behavior detection method in video based on target positioning and characteristic fusion
CN106979980A (en) * 2017-04-28 2017-07-25 武汉理工大学 A kind of new multiple spectra acoustics camera and recognition methods
CN107122712A (en) * 2017-03-27 2017-09-01 大连大学 It polymerize the palmprint image recognition methods of description vectors based on convolutional neural networks and two-way local feature
CN107133354A (en) * 2017-05-25 2017-09-05 北京小米移动软件有限公司 The acquisition methods and device of description information of image
CN107273835A (en) * 2017-06-07 2017-10-20 南京航空航天大学 Act of violence intelligent detecting method based on video analysis
CN107679522A (en) * 2017-10-31 2018-02-09 内江师范学院 Action identification method based on multithread LSTM
CN108009486A (en) * 2017-11-28 2018-05-08 中科亿海微电子科技（苏州）有限公司 Video frequency identifying method and system based on deep learning training reconstruct image
CN108021889A (en) * 2017-12-05 2018-05-11 重庆邮电大学 A kind of binary channels infrared behavior recognition methods based on posture shape and movable information
WO2018086513A1 (en) * 2016-11-08 2018-05-17 杭州海康威视数字技术股份有限公司 Target detection method and device
CN108229440A (en) * 2018-02-06 2018-06-29 北京奥开信息科技有限公司 One kind is based on Multi-sensor Fusion indoor human body gesture recognition method
CN108229580A (en) * 2018-01-26 2018-06-29 浙江大学 Sugared net ranking of features device in a kind of eyeground figure based on attention mechanism and Fusion Features
CN108288035A (en) * 2018-01-11 2018-07-17 华南理工大学 The human motion recognition method of multichannel image Fusion Features based on deep learning
CN108335402A (en) * 2017-01-18 2018-07-27 武汉卓目科技有限公司 A kind of cash inspecting machine infrared tube false distinguishing method based on deep learning
CN108596082A (en) * 2018-04-20 2018-09-28 重庆邮电大学 Human face in-vivo detection method based on image diffusion velocity model and color character
CN108664922A (en) * 2018-05-10 2018-10-16 东华大学 A kind of infrared video Human bodys' response method based on personal safety
CN108830308A (en) * 2018-05-31 2018-11-16 西安电子科技大学 A kind of Modulation Identification method that traditional characteristic signal-based is merged with depth characteristic
CN108830330A (en) * 2018-06-22 2018-11-16 西安电子科技大学 Classification of Multispectral Images method based on self-adaptive features fusion residual error net
CN108875756A (en) * 2017-05-08 2018-11-23 上海荆虹电子科技有限公司 A kind of the behavior classification acquisition methods and device of video
CN109684933A (en) * 2018-11-30 2019-04-26 广州大学 A kind of front pedestrian scurries out the method for early warning of road
CN109815967A (en) * 2019-02-28 2019-05-28 北京环境特性研究所 CNN ship seakeeping system and method based on Fusion Features
CN109977846A (en) * 2019-03-22 2019-07-05 中国科学院重庆绿色智能技术研究院 A kind of in-vivo detection method and system based on the camera shooting of near-infrared monocular
CN110135386A (en) * 2019-05-24 2019-08-16 长沙学院 A human action recognition method and system based on deep learning
CN110135352A (en) * 2019-05-16 2019-08-16 南京砺剑光电技术研究院有限公司 A kind of tactical operation appraisal procedure based on deep learning
CN110236560A (en) * 2019-06-06 2019-09-17 深圳市联邦佳维工业装备有限公司 Six axis attitude detecting methods of intelligent wearable device, system
CN110378254A (en) * 2019-07-03 2019-10-25 中科软科技股份有限公司 Recognition methods, system, electronic equipment and the storage medium of vehicle damage amending image trace
CN110555387A (en) * 2019-08-02 2019-12-10 华侨大学 Behavior identification method based on local joint point track space-time volume in skeleton sequence
CN110674884A (en) * 2019-09-30 2020-01-10 山东浪潮人工智能研究院有限公司 Image identification method based on feature fusion
US10600208B2 (en) 2017-12-21 2020-03-24 Industrial Technology Research Institute Object detecting device, object detecting method and non-transitory computer-readable medium
CN110959160A (en) * 2017-08-01 2020-04-03 华为技术有限公司 A gesture recognition method, device and device
CN111079655A (en) * 2019-12-18 2020-04-28 航天物联网技术有限公司 Method for recognizing human body behaviors in video based on fusion neural network
CN111145222A (en) * 2019-12-30 2020-05-12 浙江中创天成科技有限公司 Fire detection method combining smoke movement trend and textural features
CN111310519A (en) * 2018-12-11 2020-06-19 成都智叟智能科技有限公司 Goods deep learning training method based on machine vision and data sampling
CN111400440A (en) * 2020-02-28 2020-07-10 深圳市华海同创科技有限公司 Intention identification method and device
CN111680543A (en) * 2020-04-23 2020-09-18 北京迈格威科技有限公司 Action recognition method and device and electronic equipment
CN111860086A (en) * 2019-06-26 2020-10-30 广州凡拓数字创意科技股份有限公司 Gesture recognition method, device and system based on deep neural network
CN112329815A (en) * 2020-09-30 2021-02-05 华南师范大学 Model training method, device and medium for detecting travel track abnormality
CN112597975A (en) * 2021-02-26 2021-04-02 上海闪马智能科技有限公司 Fire smoke and projectile detection method and system based on video
CN112669320A (en) * 2021-03-22 2021-04-16 四川大学 SPECT thyroid imaging intelligent identification method based on deep neural network
CN112749672A (en) * 2021-01-19 2021-05-04 携程旅游网络技术（上海）有限公司 Photo album video identification method, system, equipment and storage medium
CN112801042A (en) * 2021-03-08 2021-05-14 南京大学 Student class-taking behavior identification method based on deep learning and video tracks
CN112930535A (en) * 2019-10-07 2021-06-08 华为技术有限公司 Anomaly detection of crowd behavior based on video analysis
CN113139518A (en) * 2021-05-14 2021-07-20 杭州旭颜科技有限公司 Section bar cutting state monitoring method based on industrial internet
CN113395138A (en) * 2021-06-15 2021-09-14 重庆邮电大学 PC-SCMA joint iterative detection decoding method based on deep learning
CN109740670B (en) * 2019-01-02 2022-01-11 京东方科技集团股份有限公司 Video classification method and device
CN115147864A (en) * 2022-05-31 2022-10-04 合肥学院 Infrared human body behavior identification method based on collaborative heterogeneous deep learning network
CN115965553A (en) * 2023-01-03 2023-04-14 无锡学院 Face image local enhancement method based on multi-feature fusion
CN117528085A (en) * 2024-01-08 2024-02-06 中国矿业大学 Video compression coding method based on intelligent feature clustering
CN118587635A (en) * 2024-08-05 2024-09-03 深圳市旭景数字技术有限公司 Adaptive video content security monitoring method and system based on deep learning

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106407889B (en) * 2016-08-26 2020-08-04 上海交通大学 Method for recognizing human body interaction in video based on optical flow graph deep learning model

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8345984B2 (en) * 2010-01-28 2013-01-01 Nec Laboratories America, Inc. 3D convolutional neural networks for automatic human action recognition
CN104281853A (en) * 2014-09-02 2015-01-14 电子科技大学 Behavior identification method based on 3D convolution neural network
US20150117760A1 (en) * 2013-10-30 2015-04-30 Nec Laboratories America, Inc. Regionlets with Shift Invariant Neural Patterns for Object Detection
CN104933417A (en) * 2015-06-26 2015-09-23 苏州大学 Behavior recognition method based on sparse spatial-temporal characteristics

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8345984B2 (en) * 2010-01-28 2013-01-01 Nec Laboratories America, Inc. 3D convolutional neural networks for automatic human action recognition
US20150117760A1 (en) * 2013-10-30 2015-04-30 Nec Laboratories America, Inc. Regionlets with Shift Invariant Neural Patterns for Object Detection
CN104281853A (en) * 2014-09-02 2015-01-14 电子科技大学 Behavior identification method based on 3D convolution neural network
CN104933417A (en) * 2015-06-26 2015-09-23 苏州大学 Behavior recognition method based on sparse spatial-temporal characteristics

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
雷庆等: "复杂场景下的人体行为识别研究新进展", 《计算机科学》 *

Cited By (73)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106446937A (en) * 2016-09-08 2017-02-22 天津大学 Multi-convolution identifying system for AER image sensor
CN106355210B (en) * 2016-09-14 2019-03-19 华北电力大学（保定） Insulator Infrared Image feature representation method based on depth neuron response modes
CN106355210A (en) * 2016-09-14 2017-01-25 华北电力大学（保定） Method for expressing infrared image features of insulators on basis of depth neuron response modes
WO2018086513A1 (en) * 2016-11-08 2018-05-17 杭州海康威视数字技术股份有限公司 Target detection method and device
US10949673B2 (en) * 2016-11-08 2021-03-16 Hangzhou Hikvision Digital Technology Co., Ltd. Target detection method and device
CN106709447A (en) * 2016-12-21 2017-05-24 华南理工大学 Abnormal behavior detection method in video based on target positioning and characteristic fusion
CN108335402A (en) * 2017-01-18 2018-07-27 武汉卓目科技有限公司 A kind of cash inspecting machine infrared tube false distinguishing method based on deep learning
CN107122712A (en) * 2017-03-27 2017-09-01 大连大学 It polymerize the palmprint image recognition methods of description vectors based on convolutional neural networks and two-way local feature
CN106979980B (en) * 2017-04-28 2020-06-09 武汉理工大学 A kind of multi-spectral acoustic camera and identification method
CN106979980A (en) * 2017-04-28 2017-07-25 武汉理工大学 A kind of new multiple spectra acoustics camera and recognition methods
CN108875756A (en) * 2017-05-08 2018-11-23 上海荆虹电子科技有限公司 A kind of the behavior classification acquisition methods and device of video
CN107133354A (en) * 2017-05-25 2017-09-05 北京小米移动软件有限公司 The acquisition methods and device of description information of image
CN107273835A (en) * 2017-06-07 2017-10-20 南京航空航天大学 Act of violence intelligent detecting method based on video analysis
CN110959160A (en) * 2017-08-01 2020-04-03 华为技术有限公司 A gesture recognition method, device and device
CN110959160B (en) * 2017-08-01 2024-12-20 华为技术有限公司 A gesture recognition method, device and equipment
CN107679522B (en) * 2017-10-31 2020-10-13 内江师范学院 Multi-stream LSTM-based action identification method
CN107679522A (en) * 2017-10-31 2018-02-09 内江师范学院 Action identification method based on multithread LSTM
CN108009486A (en) * 2017-11-28 2018-05-08 中科亿海微电子科技（苏州）有限公司 Video frequency identifying method and system based on deep learning training reconstruct image
CN108009486B (en) * 2017-11-28 2021-06-08 中科亿海微电子科技（苏州）有限公司 Video identification method and system based on deep learning training reconfiguration picture
CN108021889A (en) * 2017-12-05 2018-05-11 重庆邮电大学 A kind of binary channels infrared behavior recognition methods based on posture shape and movable information
US10600208B2 (en) 2017-12-21 2020-03-24 Industrial Technology Research Institute Object detecting device, object detecting method and non-transitory computer-readable medium
CN108288035A (en) * 2018-01-11 2018-07-17 华南理工大学 The human motion recognition method of multichannel image Fusion Features based on deep learning
CN108229580B (en) * 2018-01-26 2020-12-11 浙江大学 A device for classification of sugar network features in fundus images based on attention mechanism and feature fusion
CN108229580A (en) * 2018-01-26 2018-06-29 浙江大学 Sugared net ranking of features device in a kind of eyeground figure based on attention mechanism and Fusion Features
CN108229440A (en) * 2018-02-06 2018-06-29 北京奥开信息科技有限公司 One kind is based on Multi-sensor Fusion indoor human body gesture recognition method
CN108596082A (en) * 2018-04-20 2018-09-28 重庆邮电大学 Human face in-vivo detection method based on image diffusion velocity model and color character
CN108664922A (en) * 2018-05-10 2018-10-16 东华大学 A kind of infrared video Human bodys' response method based on personal safety
CN108830308B (en) * 2018-05-31 2021-12-14 西安电子科技大学 A modulation recognition method based on the fusion of traditional features and deep features of signals
CN108830308A (en) * 2018-05-31 2018-11-16 西安电子科技大学 A kind of Modulation Identification method that traditional characteristic signal-based is merged with depth characteristic
CN108830330B (en) * 2018-06-22 2021-11-02 西安电子科技大学 Multispectral image classification method based on adaptive feature fusion residual network
CN108830330A (en) * 2018-06-22 2018-11-16 西安电子科技大学 Classification of Multispectral Images method based on self-adaptive features fusion residual error net
CN109684933A (en) * 2018-11-30 2019-04-26 广州大学 A kind of front pedestrian scurries out the method for early warning of road
CN111310519B (en) * 2018-12-11 2024-01-05 成都智叟智能科技有限公司 Goods deep learning training method based on machine vision and data sampling
CN111310519A (en) * 2018-12-11 2020-06-19 成都智叟智能科技有限公司 Goods deep learning training method based on machine vision and data sampling
CN109740670B (en) * 2019-01-02 2022-01-11 京东方科技集团股份有限公司 Video classification method and device
CN109815967A (en) * 2019-02-28 2019-05-28 北京环境特性研究所 CNN ship seakeeping system and method based on Fusion Features
CN109977846A (en) * 2019-03-22 2019-07-05 中国科学院重庆绿色智能技术研究院 A kind of in-vivo detection method and system based on the camera shooting of near-infrared monocular
CN110135352B (en) * 2019-05-16 2023-05-12 南京砺剑光电技术研究院有限公司 Tactical action evaluation method based on deep learning
CN110135352A (en) * 2019-05-16 2019-08-16 南京砺剑光电技术研究院有限公司 A kind of tactical operation appraisal procedure based on deep learning
CN110135386A (en) * 2019-05-24 2019-08-16 长沙学院 A human action recognition method and system based on deep learning
CN110236560A (en) * 2019-06-06 2019-09-17 深圳市联邦佳维工业装备有限公司 Six axis attitude detecting methods of intelligent wearable device, system
CN111860086A (en) * 2019-06-26 2020-10-30 广州凡拓数字创意科技股份有限公司 Gesture recognition method, device and system based on deep neural network
CN110378254B (en) * 2019-07-03 2022-04-19 中科软科技股份有限公司 Method and system for identifying vehicle damage image modification trace, electronic device and storage medium
CN110378254A (en) * 2019-07-03 2019-10-25 中科软科技股份有限公司 Recognition methods, system, electronic equipment and the storage medium of vehicle damage amending image trace
CN110555387A (en) * 2019-08-02 2019-12-10 华侨大学 Behavior identification method based on local joint point track space-time volume in skeleton sequence
CN110555387B (en) * 2019-08-02 2022-07-19 华侨大学 Behavior identification method based on space-time volume of local joint point track in skeleton sequence
CN110674884A (en) * 2019-09-30 2020-01-10 山东浪潮人工智能研究院有限公司 Image identification method based on feature fusion
CN112930535A (en) * 2019-10-07 2021-06-08 华为技术有限公司 Anomaly detection of crowd behavior based on video analysis
CN111079655A (en) * 2019-12-18 2020-04-28 航天物联网技术有限公司 Method for recognizing human body behaviors in video based on fusion neural network
CN111145222A (en) * 2019-12-30 2020-05-12 浙江中创天成科技有限公司 Fire detection method combining smoke movement trend and textural features
CN111400440A (en) * 2020-02-28 2020-07-10 深圳市华海同创科技有限公司 Intention identification method and device
CN111680543A (en) * 2020-04-23 2020-09-18 北京迈格威科技有限公司 Action recognition method and device and electronic equipment
CN111680543B (en) * 2020-04-23 2023-08-29 北京迈格威科技有限公司 Action recognition method and device and electronic equipment
CN112329815B (en) * 2020-09-30 2022-07-22 华南师范大学 Model training method, driving trajectory abnormality detection method, device and medium
CN112329815A (en) * 2020-09-30 2021-02-05 华南师范大学 Model training method, device and medium for detecting travel track abnormality
CN112749672A (en) * 2021-01-19 2021-05-04 携程旅游网络技术（上海）有限公司 Photo album video identification method, system, equipment and storage medium
CN112597975A (en) * 2021-02-26 2021-04-02 上海闪马智能科技有限公司 Fire smoke and projectile detection method and system based on video
CN112801042B (en) * 2021-03-08 2023-12-15 南京大学 Student lesson behavior recognition method based on deep learning and video track
CN112801042A (en) * 2021-03-08 2021-05-14 南京大学 Student class-taking behavior identification method based on deep learning and video tracks
CN112669320A (en) * 2021-03-22 2021-04-16 四川大学 SPECT thyroid imaging intelligent identification method based on deep neural network
CN112669320B (en) * 2021-03-22 2021-08-13 四川大学 An intelligent identification method of SPECT thyroid imaging based on deep neural network
CN113139518B (en) * 2021-05-14 2022-07-29 江苏中天互联科技有限公司 Section bar cutting state monitoring method based on industrial internet
CN113139518A (en) * 2021-05-14 2021-07-20 杭州旭颜科技有限公司 Section bar cutting state monitoring method based on industrial internet
CN113395138B (en) * 2021-06-15 2022-05-03 重庆邮电大学 PC-SCMA joint iterative detection and decoding method based on deep learning
CN113395138A (en) * 2021-06-15 2021-09-14 重庆邮电大学 PC-SCMA joint iterative detection decoding method based on deep learning
CN115147864A (en) * 2022-05-31 2022-10-04 合肥学院 Infrared human body behavior identification method based on collaborative heterogeneous deep learning network
CN115147864B (en) * 2022-05-31 2024-08-30 合肥学院 Infrared human behavior recognition method based on cooperative heterogeneous deep learning network
CN115965553A (en) * 2023-01-03 2023-04-14 无锡学院 Face image local enhancement method based on multi-feature fusion
CN115965553B (en) * 2023-01-03 2023-12-19 无锡学院 Multi-feature fusion-based face image local enhancement method
CN117528085A (en) * 2024-01-08 2024-02-06 中国矿业大学 Video compression coding method based on intelligent feature clustering
CN117528085B (en) * 2024-01-08 2024-03-19 中国矿业大学 A video compression coding method based on intelligent feature clustering
CN118587635A (en) * 2024-08-05 2024-09-03 深圳市旭景数字技术有限公司 Adaptive video content security monitoring method and system based on deep learning
CN118587635B (en) * 2024-08-05 2024-10-29 深圳市旭景数字技术有限公司 Adaptive video content security monitoring method and system based on deep learning

Also Published As

Publication number Publication date
CN105787458B (en) 2019-01-04

Similar Documents

Publication Publication Date Title
CN105787458B (en) The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature
CN113158875B (en) Image-text sentiment analysis method and system based on multi-modal interaction fusion network
CN110334705B (en) A language recognition method for scene text images combining global and local information
CN110111335B (en) Urban traffic scene semantic segmentation method and system for adaptive countermeasure learning
CN109446923B (en) Deeply supervised convolutional neural network behavior recognition method based on training feature fusion
CN108596958B (en) A Target Tracking Method Based on Difficult Positive Sample Generation
CN108648746B (en) A method for generating open-domain video natural language description based on multimodal feature fusion
CN113158862B (en) A lightweight real-time face detection method based on multi-task
CN110929679B (en) GAN-based unsupervised self-adaptive pedestrian re-identification method
CN110097000A (en) Video behavior recognition methods based on local feature Aggregation Descriptor and sequential relationship network
CN112766334B (en) A cross-domain image classification method based on pseudo-label domain adaptation
CN108681712A (en) A kind of Basketball Match Context event recognition methods of fusion domain knowledge and multistage depth characteristic
CN107563279A (en) The model training method adjusted for the adaptive weighting of human body attributive classification
CN108491766B (en) End-to-end crowd counting method based on depth decision forest
CN108427740B (en) Image emotion classification and retrieval algorithm based on depth metric learning
CN114049381A (en) A Siamese Cross-Target Tracking Method Fusing Multi-layer Semantic Information
CN110674483B (en) An identity recognition method based on multimodal information
CN108764019A (en) A kind of Video Events detection method based on multi-source deep learning
US11223782B2 (en) Video processing using a spectral decomposition layer
CN112801019B (en) Method and system for eliminating unsupervised vehicle re-identification bias based on synthetic data
Ocquaye et al. Dual exclusive attentive transfer for unsupervised deep convolutional domain adaptation in speech emotion recognition
CN116342942A (en) Cross-domain target detection method based on multistage domain adaptation weak supervision learning
CN117523614A (en) A YOLOv7 head recognition method based on fused NAM
CN116433721A (en) An Outdoor RGB-T Target Tracking Algorithm Based on Generated Pseudo Fusion Features
Qin et al. Application of video scene semantic recognition technology in smart video

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant